US stocks Micron rose 5.42%, SanDisk rose 6.66%; A-shares Biwin Storage rose 8.36%, Puya Semiconductor rose 11.38%; Montage Technology, which makes memory interfaces, rose only 6.20% — all benefit from memory price increases, so why did interface chips rise the least?
- So its gross margin moves up along the 68% line, not Biwin's jump from 14% to 51%.
- Interconnect chips 20.47 yi, year-on-year up 49.3%, quarter-on-quarter up 20.9%
- This part's gross margin 68.0%~68.5%, year-on-year up only 2.31~2.81 percentage points.
- Net profit 16.02 yi, year-on-year up 525.34%
- Puya rose 11.38%, and part of that is not about memory.
US stocks Micron rose 5.42%, SanDisk rose 6.66%; A-shares Biwin Storage rose 8.36%, Puya Semiconductor rose 11.38%; Montage Technology, which makes memory interfaces, rose only 6.20% — all benefit from memory price increases, so why did interface chips rise the least?
美股美光涨5.42%、闪迪涨6.66%,A股佰维存储涨8.36%、普冉股份涨11.38%,做内存接口的澜起科技只涨6.20%——同样因存储涨价受益,为什么接口芯片涨得最少?
The desk answers已核实Walnut, from public filings
Micron, SanDisk, Biwin, Puya, and Montage—these five names hang under the same "memory price increase" story, but the money from the price increase only lands in two places: those that can price memory themselves, and those holding a pile of memory in hand. Montage occupies neither—what it sells is interface chips on memory modules, counted one by one per module, and how much memory dies rise in price has nothing to do with the price of its logic chip. That is why it rises the least—not because it was left behind, but because it is not on that price chain.
The receiving end is Micron. The 8-K of September 30: quarterly revenue 542.3 yi dollars, gross margin 86.8%, operating margin 80.7%; a year earlier in the same period gross margin was 44.7%. These forty-two percentage points were not saved—they were sold. Next-quarter guidance 615 yi dollars, gross margin 85.95%. SanDisk stands on the same node—a NAND original manufacturer, the price increases fall directly into its own gross margin. This is where the two words "price increase" land.
The A-share company most like Micron is Biwin Storage. Today's third-quarter report: quarterly revenue 61.42 yi, year-on-year up 130.67%; net profit 16.02 yi, year-on-year up 525.34%; excluding non-recurring items 18.34 yi, year-on-year up 759.88%. Two more lines are more critical: first three quarters revenue 217.17 yi, operating cost 105.51 yi, and the gross margin calculated by subtraction is about 51.4%; last year same period 65.75 yi versus 56.60 yi, about 13.9%. Inventory rose from 78.68 yi at the start of the year to 240.53 yi—goods pressed in hand, cost still at earlier low prices, selling at risen prices, gross margin lifted from 14% to 51%, earning from inventory and price spread.
Puya rose 11.38%, and part of that is not about memory. On October 14 the Shanghai Stock Exchange M&A and Reorganization Committee reviewed its plan to acquire 49% equity of Zhuhai Noah Changtian, announced October 15, still awaiting CSRC registration. This is the company's own matter.
Montage's October 21 third-quarter earnings pre-increase announcement: quarterly revenue 20.79 yi, year-on-year up 46.0%; interconnect chips 20.47 yi, year-on-year up 49.3%, quarter-on-quarter up 20.9%; this part's gross margin 68.0%~68.5%, year-on-year up only 2.31~2.81 percentage points. The company itself wrote the growth reason as server CPU demand driving up demand for memory modules and interconnect chips, and rising shipment share of DDR5 third-to-fifth-subgeneration RCD—all volume and mix, not a single word about price.
The core is here: Montage sells by the piece. On one server memory module it sells RCD, sells matching SPD and PMIC, and on MRDIMM also MRCD/MDB; how many pieces and what to configure is set by the JEDEC standard, not by memory price. Whether the die goes from 100 kuai to 1000 kuai, those few interface chips on the module stay at that price. Where it can raise prices is generations—subgeneration 3 to 4 to 5, the new generation a bit more expensive, one small step a year. So its gross margin moves up along the 68% line, not Biwin's jump from 14% to 51%.
There is another layer: it was already at 68%. Price increases give it no leverage, and room to go higher is limited; Biwin goes up from 14%, and every cent of price increase is almost pure profit. The market chases this elasticity. In the first half what ran faster for it were PCIe Retimer and CXL, totaling 5.38 yi, year-on-year up 80.7%—its growth story is in AI interconnect, not in the die cycle.
The strongest opposing argument: "price increases do not pass through Montage" is not entirely right. It has been raising prices all along, only in generations—first-half interconnect chip gross margin 69.3%, year-on-year up 4.95 percentage points. One round up in subgeneration, and unit price and gross margin lift a notch; this channel truly exists. If in 2027 subgeneration 5 and MRCD/MDB ramp up, the channel will steepen, and then today's 6.20% is mispriced: the market treats it as a component maker that does not raise prices, when in fact it quietly raises prices every generation.
What would count as me being wrong: in Montage's 2026 annual report, if interconnect chip gross margin rises another 3 percentage points or more above the third quarter's 68.0%~68.5%, and the company writes the reason as price rather than subgeneration mix, then my "price increases do not pass through it" does not hold; or if fourth-quarter interconnect chip revenue no longer grows quarter-on-quarter (below 20.47 yi), it means its small rise is not insensitivity to price but volume loosening first.
What is really moving today is the other two nodes. Lam Research's October 21 8-K: quarterly revenue 82.2 yi dollars, gross margin 52.3%, operating margin 40.0%, CEO says AI is driving rising semiconductor demand and increasing device complexity. Fabrinet's August 17 8-K: quarterly revenue 13.16 yi dollars, year-on-year up 45%, gross margin only 12.0%—also called AI hardware, contract assembly takes 12%, the original manufacturer takes 86.8%, and under "beneficiary" lies a gap of more than seventy percentage points in pricing power.
Back to the question: it is not that interface chips are not favored, it is that they stand beside the price chain. On this chain money stops at the two nodes of "can price" and "has inventory"; Montage eats from volume and generation shifts. It rises 6.20% today, and what rises is volume, not price. Explaining it with "price increase" is the most convoluted; explaining it with "server shipments plus DDR5 generation switching" is what matches the statements.
Sources: Micron 8-K (2026-09-30, EX-99.1, 0000723125-26-000018); Lam Research 8-K (2026-10-21, EX-99.1, 0000707549-26-000052); Fabrinet 8-K (2026-08-17, EX-99.1, 0001408710-26-000026); Montage Technology "Voluntary Disclosure Announcement of Expected Increase in Third Quarter 2026 Results" (2026-10-21, Announcement No. 2026-057, Cninfo 1225612915); Montage Technology "2026 Semi-Annual Report" (2026-08-29, Cninfo 1225530486); Biwin Storage "2026 Third Quarter Report" (2026-10-29, Cninfo 1225638424); Puya Semiconductor "Announcement on the Approval by the Merger, Acquisition and Reorganization Review Committee of the Shanghai Stock Exchange of the Matter of Acquiring Assets by Issuing Shares, Convertible Corporate Bonds and Paying Cash and Concurrently Raising Supporting Funds" (2026-10-15, Announcement No. 2026-091, Cninfo 1225605603).
美光、闪迪、佰维、普冉、澜起这五个名字挂在同一个"存储涨价"故事下,但涨价的钱只落在两个地方:能自己给内存定价的,和手里压着一堆内存的。澜起两个都不占——它卖的是内存条上的接口芯片,一颗一颗按根条子算,内存颗粒涨多少跟它那颗逻辑芯片的价钱没关系。所以它涨得最少,不是被落下,是它不在那条价格链上。
收钱的那一头看美光。9月30日的8-K:单季收入542.3亿美元,毛利率86.8%,营业利润率80.7%;一年前同期毛利率44.7%。这四十二个百分点不是省出来的,是卖出来的。下一季指引615亿美元、毛利率85.95%。闪迪跟它站在同一节——NAND原厂,涨的价直接落进自己的毛利。这就是"涨价"两个字的落点。
A股最像美光的是佰维存储。今天的三季报:单季收入61.42亿,同比增130.67%;净利16.02亿,同比增525.34%;扣非18.34亿,同比增759.88%。两行数更关键:前三季度收入217.17亿、营业成本105.51亿,相减算出的毛利率约51.4%;去年同期65.75亿对56.60亿,约13.9%。存货从年初78.68亿增到240.53亿——货压在手里,成本还是早先的低价,卖的是涨完的价,毛利从14%抬到51%,赚的是库存和价差。
普冉涨11.38%,有一部分不是存储的事。10月14日上交所并购重组委审过它收购珠海诺亚长天49%股权的方案,10月15日公告,还要等证监会注册。这是公司自己的事。
澜起10月21日的三季度业绩预增公告:单季收入20.79亿,同比增46.0%;互连类芯片20.47亿,同比增49.3%,环比增20.9%;这部分毛利率68.0%~68.5%,同比只提高2.31~2.81个百分点。公司自己写的增长原因是服务器CPU需求带动内存模组和互连芯片需求量上行、DDR5第三到第五子代RCD出货占比提升——全是量和结构,一个字没提价格。
核心在这儿:澜起按颗卖。一根服务器内存条上它卖RCD、卖配套的SPD和PMIC,MRDIMM上还要MRCD/MDB;配几颗、配什么由JEDEC标准规定,不由内存价格定。颗粒从100块涨到1000块,条子上那几颗接口芯片还是那个价。它能提价的地方是代际,子代3换4换5,新一代贵一点,一年走一小步。所以它的毛利率是68%这条线往上挪,不是佰维那种从14%跳到51%。
还有一层:它本来就在68%上。涨价对它没有杠杆,再往上的空间有限;佰维从14%往上,每一分涨价几乎都是纯利润。市场追的是这个弹性。上半年它跑得更快的是PCIe Retimer和CXL那几块,合计5.38亿、同比增80.7%——它的成长故事在AI互连,不在颗粒行情。
反方最强的一条:"涨价不经过澜起"不完全对。它一直在提价,只是提在代际上——上半年互连类芯片毛利率69.3%,同比提高4.95个百分点。子代往上走一轮,单价和毛利率就抬一档,这条通道真实存在。若2027年子代5和MRCD/MDB放量,通道会变陡,那时今天这6.20%就是定错了价:市场把它当成一个不涨价的零件厂,它其实每代都在悄悄提价。
什么情况算我看错:澜起2026年报里,互连类芯片毛利率若比三季度的68.0%~68.5%再抬3个百分点以上,且公司把原因写成价格而不是子代占比,那我"涨价不经过它"就站不住;或者四季度互连类芯片收入环比不再增长(低于20.47亿),说明它涨得少不是对价格不敏感,是量先松了。
今天真正在动的是另外两节。泛林10月21日的8-K:单季收入82.2亿美元,毛利率52.3%,营业利润率40.0%,CEO说AI带动半导体需求上升、器件复杂度提高。Fabrinet 8月17日的8-K:单季收入13.16亿美元,同比增45%,毛利率只有12.0%——同样叫AI硬件,代工组装拿12%,原厂拿86.8%,"受益"底下差着七十多个百分点的定价权。
回到问题:不是接口芯片不被看好,是它站在价格链旁边。这条链上钱停在"能定价"和"有存货"两节,澜起靠量和换代吃饭。它今天涨6.20%,涨的是量不是价。用"涨价"解释它最绕,用"服务器出货加DDR5代际切换"解释它,才对得上报表。
来源:美光8-K(2026-09-30,EX-99.1,0000723125-26-000018);泛林8-K(2026-10-21,EX-99.1,0000707549-26-000052);Fabrinet 8-K(2026-08-17,EX-99.1,0001408710-26-000026);澜起科技《2026年第三季度业绩预增的自愿性披露公告》(2026-10-21,公告编号2026-057,巨潮1225612915);澜起科技《2026年半年度报告》(2026-08-29,巨潮1225530486);佰维存储《2026年第三季度报告》(2026-10-29,巨潮1225638424);普冉股份《关于发行股份、可转换公司债券及支付现金购买资产并募集配套资金事项获得上海证券交易所并购重组审核委员会审核通过的公告》(2026-10-15,公告编号2026-091,巨潮1225605603)。
先把这一天放回它所在的一个月。过去20个交易日,这六只票还是跌了21%到32%——这个数已经算进了10月28日的反弹。同一个板块同涨同跌的一个月里,拿一天排「谁更受益」,信息量很低。
第一层是时点。澜起的三季报预告10月21日就出了,当天涨5.36%,10月26日又跌6.96%,好消息已经用掉;佰维的三季报10月29日才披露,10月28日还在报表之前;德明利的三季报10月26日已经出了,毛利率回落到32.89%,10月28日它只涨2.93%。这一天的排序,有很大一部分是「谁的报表还没公开」。
| 公司 | 今日(%) | 近20日(%) | 成交额(亿) | PE(滚动) |
|---|---|---|---|---|
| 澜起科技 | 6.20 | (21.2) | 57.2 | 52* |
| 佰维存储 | 8.36 | (24.5) | 45.9 | 8.1 |
| 普冉股份 | 11.38 | (31.8) | 33.3 | 43.7 |
| 德明利 | 2.93 | (21.2) | 53.2 | 5.7 |
| 聚辰股份 | 7.76 | (31.6) | 9.2 | 35.2 |
| 君正股份 | 18.37 | (23.4) | 72.7 | 17.3 |
*澜起按Q3预告净利中值12.5亿重算,滚动12个月净利38.5亿;其余按最近四个已披露季度。
第二层是盘子。澜起这一天成交57.2亿,比普冉的33.3亿多72%,但换手率只有2.88%,普冉是7.87%——它的流通市值1985亿,普冉423亿。同样一笔钱,推进2000亿的票和推进400亿的票,涨幅不可能一样。君正换手15.71%,是这天最高的。
## 估值早就把两类生意分好了
第三层是估值。佰维8.1倍、德明利5.7倍,澜起按Q3预告算是52倍,聚辰35.2倍、普冉43.7倍。市场给模组厂的是「周期顶利润」的倍数,给澜起的是「能穿过周期」的倍数。
同样一条涨价消息,对8倍的模组厂是单季利润几倍的增长,对52倍的澜起只是毛利率挪半个点的事。前者的股价反应自然更大。澜起涨得少,有一部分是它早就被付过钱了。
## 真正在分化的是毛利率的方向
如果排序全是噪音,不会刚好把君正排在第一、澜起排在最后。真正的信息在毛利率的环比方向。
| 公司 | 二季度单季毛利率(%) | 三季度单季毛利率(%) |
|---|---|---|
| 君正股份 | 57.12 | 65.93 |
| 聚辰股份 | 53.72 | 56.77 |
| 澜起科技(互连类) | 67.40 | 68.0~68.5 |
| 佰维存储 | 55.98 | 42.81 |
| 德明利 | 38.84 | 32.89 |
同行的单季毛利率摆在一起,方向就清楚了:君正和聚辰在往上走,澜起几乎是一条平线,佰维和德明利在往下掉。这一天涨18.37%的君正,管理层10月27日的调研记录里说得很直白:
> 四季度存储价格整体呈上涨趋势,NOR Flash大容量产品持续涨价……预计四季度存储毛利率会继续增长。
澜起的预告里没有这句话。它写的是:
> DDR5内存接口及模组配套芯片出货量大幅增加,DDR5第三、第四、第五子代RCD芯片出货占比进一步提升。
一个是价,一个是量。这一天的涨幅榜,基本上就是这两句话的排序。普冉三季报还没出,它的11.38%不能当毛利率的证据。
## 涨价的账落在营业成本
这条链的卡点在颗粒:全球DRAM九成左右的份额在三星、海力士、美光三家手里,价格由它们定。模组厂的毛利率为什么会掉?看成本端。德明利第三季度营业成本环比增36.3%,收入环比增24.3%;佰维营业成本环比降8.9%,收入环比降29.9%。这两个成本数是按「收入×(1-毛利率)」倒算的。成本降得比收入慢,毛利率就一定掉。
传导链是:颗粒涨价 → 模组厂营业成本 → 单季毛利率。它先落在「营业成本」这个科目,不是「税金及附加」,也不是「资产减值」。模组厂先吃低价库存的红利,等库存用完,新补的料是涨价后的成本,毛利率就往回走。
```chart
type: line
title: 模组厂的单季毛利率在二季度见顶,三季度已经回落
unit: %
labels: 2025Q4, 2026Q1, 2026Q2, 2026Q3
values: 31.91, 53.30, 55.98, 42.81
values2: 27.14, 57.42, 38.84, 32.89
legend: 佰维|德明利
```
佰维Q3还提了0.97亿存货跌价。240亿存货不是纯资产,里面有跌价的部分。
利润是真的,但它长在资产负债表上:存货240.53亿(年初78.68亿)、预付款项73.75亿(年初1.55亿)、短期借款107.19亿(年初42.74亿)、经营性现金流-100.41亿。公司自己的解释是:
> 主要系公司经营性采购支出增加所致。
赚了87.68亿净利,经营性现金流是-100.41亿。这不是造假,是生意的形状——模组厂在涨价周期里,利润和现金是反向的。
## 澜起收的是「条」的钱
澜起不一样的地方,不在「接口」这两个字,在它的收费单位。
一根DDR5服务器内存条,按JEDEC的定义要配一颗RCD、一颗SPD、一颗PMIC、两颗TS;MRDIMM是「1+10」,一颗MRCD加十颗MDB。它收的是「每根条子配几颗」的钱,不是「每GB多少钱」的钱。它受益的是服务器出货量和DDR5渗透率,不是颗粒价格。
颗粒从100块涨到1000块,这根条子上那几颗逻辑芯片的单价不会跟着涨十倍。颗粒涨价改变的是每根条子的容量,不改变条数——除非客户为了压BOM少插通道,那才是澜起真正的风险。
它的毛利率只能靠代际往上挪。子代3换4换5,新产品占比提高,毛利率走一档。Q3互连类毛利率68.0%~68.5%,同比提高2.31~2.81个点,环比只提高0.57~1.07个点。这是年度台阶,不是季度斜率。
还有一个坑:按预告算,Q3整体毛利率从二季度的61.8%跳到67%出头,别把它当成提价。二季度低毛利的津逮收入是1.81亿,三季度只剩0.32亿,是结构混出来的。干净的线是互连类的67.4%到68.0%~68.5%。
## 同一条内存条上的对照
聚辰也在这条条子上,做DDR5 SPD。它第三季度收入4.51亿,环比增40.0%;扣非净利1.40亿,环比增110.87%;单季毛利率56.77%,比二季度的53.72%高3.05个点。
它这一天涨7.76%,只比澜起多1.56个点。同一条条子、同一批客户、同一个季度,一颗毛利率在升、一颗只挪了不到一个点,涨幅差1.56个点。所以「接口芯片被歧视」这个说法站不住——市场没有按「接口」和「存储」分,市场按「毛利率还在不在往上走」分。
聚辰Q3净利-0.39亿,是盛合晶微战配公允价值变动亏2.10亿,不影响毛利率这条线。
## 反方最强的一条
说澜起「没有价格敞口」,最强的反驳是:它的毛利率一直在往上走。Q3扣非净利9.5亿到10.5亿,同比增152.6%到179.2%,这个增速不比任何人慢;互连类毛利率同比提高2.31~2.81个点,公司把原因写成「随着DDR5新子代RCD芯片及互连类芯片新产品收入占比提高」。如果新一代不是卖得更贵,毛利率不会跟着代际走——所以代际本身就是一种提价。
再往远看,MRDIMM的「1+10」是真正的期权:一根条子从一颗RCD变成一颗MRCD加十颗MDB,单条要配的接口芯片从1颗变成11颗。如果2027年MRDIMM随着服务器新平台放量,澜起的「每条收入」会跳一档,那时候今天这6.20%就是定错了价。
这个反驳我接受一半。代际提价是真的,MRDIMM的期权也是真的。但它们都是「期权」,不是「弹性」——行权的条件在JEDEC的标准、在Intel和AMD的平台导入、在内存厂的产能,不在澜起手里。而且52倍PE里,已经付过一部分期权的钱。
## 什么情况算看错
- 如果四季度互连类收入环比不增、低于20.47亿,这条判断就错了——它涨得少不是对价格不敏感,是量先松了。
- 互连类毛利率再抬3个点以上,且公司把原因写成价格而不是子代占比,那「没有价格敞口」就站不住。
- DDR5服务器模组的条数下降,或者MRDIMM到2027年出货占比仍是个位数,量逻辑和期权同时落空。
回到问题。接口芯片涨得最少,不是涨价轮不到它,是它从来不在「按GB计价」的那张桌子上。它按条收费,按代际提价;互连类68%的毛利率,比这条链上做模组、做配套的都高,也是离颗粒价格最远的一节。
这一天它排最后,明天颗粒再涨一倍,它大概率还是排最后。这不是意外,是它的股价性格。
来源:澜起科技《2026年第三季度业绩预增的自愿性披露公告》(2026-10-21,公告编号2026-057,巨潮1225612915);澜起科技《2026年半年度报告》(2026-08-29,巨潮1225530486);佰维存储《2026年第三季度报告》(2026-10-29,巨潮1225638424);佰维存储《关于公司2026年前三季度计提资产减值准备的公告》(2026-10-29,公告编号2026-101,巨潮1225638423);聚辰股份《2026年第三季度报告》(2026-10-24,巨潮1225620951);德明利《2026年三季度报告》(2026-10-26,巨潮1225620615);君正股份《2026年三季度报告》(2026-10-28,巨潮1225631045)及《投资者关系活动记录表》(2026-10-28,编号20261027,巨潮1225637336)。